FLOPSCOPE · A FLOP-COUNTING NUMPY

Count every FLOP.

A NumPy-compatible math library that counts every FLOP analytically, so compute budgets stop being guesswork.

uv add flopscope

508Operations

NumPy-compatible operations with analytical FLOP costs.

πSymmetry-aware

Symmetry metadata propagates through the chain, so costs scale with unique elements.

∞Composable

Budget contexts, namespaces, and per-operation breakdowns.

IN ACTION

What does this code cost?

Same five-layer MLP, written twice. The one on the right counts every FLOP analytically as it runs.

NumPy

import numpy as np
 
 
depth, width = 5, 256
 
# Weight init
scale = np.sqrt(2 / width)
weights = [
    np.random.randn(width, width) * scale
    for _ in range(depth)
]
 
# Forward pass
x = np.random.randn(width)
h = x
for i, W in enumerate(weights):
    h = np.einsum('ij,j->i', W, h)
    if i < depth - 1:
        h = np.maximum(h, 0)
# Total FLOPs? No idea.

Flopscope

import flopscope as flops
import flopscope.numpy as fnp
 
depth, width = 5, 256
 
# Weight init
scale = fnp.sqrt(2 / width)
weights = [
    fnp.random.randn(width, width) * scale
    for _ in range(depth)
]
 
# Forward pass
x = fnp.random.randn(width)
h = x
for i, W in enumerate(weights):
    h = fnp.einsum('ij,j->i', W, h)
    if i < depth - 1:
        h = fnp.maximum(h, 0)
flops.budget_summary()  # 984,321 FLOPs

The answer, on the right: 984,321 FLOPs — counted analytically as the NumPy call runs.

READ ON

Explore the docs.

Get started

Install flopscope and create your first FLOP budget.

Migrate from NumPy

Convert existing code with FLOP annotations.

API reference

Search all 508 operations with interactive filters.

FLOP counting model

How costs are computed analytically.

For AI agents

Machine-readable resources and rules.